每周精选 · 第 0 期 Weekly picks · Edition 0
创刊回顾 · 10 月 14–27 日 Inaugural retrospective · October 14–27 2026-10-14 → 2026-10-27 · 15 篇方法 · 1 篇 Harness 评测与理论 15 method papers · 1 Harness evaluation and theory paper
首期汇总了 10 月 14–27 日的 Harness 自进化论文；从 10 月 28 日起，今日动态只收时间窗内的新论文。所列论文不是当日新增。
This inaugural edition reviews Harness self-evolution papers from October 14–27. From October 28, daily editions include only papers in their date window. These are not today's new publications.
1.1 搜索与演化主线 1.1 Search & evolution (frozen model) 2
原始发表 Published 2026-10-21
1.1 搜索与演化主线 1.1 Search & evolution (frozen model)
SERA
改了什么：VLM 规划器的 Harness
From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation
固定 VLA 执行器，可编辑 Harness 依执行经验自进化
The VLM planner evolves its editable Harness from execution experience while the VLA executor remains fixed.
HF 推荐： HF recommendations: 2026-10-22
✓ 有代码链接 Code link 来源 Source https://github.com/StevenKKXS/SERA ✓ 留出集 Held-out 来源 Source fixed before held-out testing 其余 2 项待核 2 other checks await verification 跨模型：待核 Cross-model: Not verified 同预算：待核 Matched budget: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-20
1.1 搜索与演化主线 1.1 Search & evolution (frozen model) 已收录证据库 In the evidence library
RiseHarness
改了什么：详见证据卡
Harness Self-Evolution for Long-Horizon Terminal Agents
针对重复运行时时成时败的任务：结合 Agent 轨迹与运行时内部追踪，用 Harness Handbook 把行为映射到源码，并检索其他框架的实现作参考。
Targets tasks with inconsistent success across repeated runs: combines agent trajectories and internal runtime traces, maps behavior to source code using a Harness Handbook, and retrieves implementations from other frameworks as references.
HF 推荐： HF recommendations: 2026-10-21
✓ 留出集 Held-out 来源 Source papers/riseharness.md 其余 3 项待核 3 other checks await verification 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 有代码链接：待核 Code link: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
1.2 泛化、正则、准入与成本 1.2 Generalization, regularization, admission & cost 1
原始发表 Published 2026-10-19
1.2 泛化、正则、准入与成本 1.2 Generalization, regularization, admission & cost 已收录证据库 In the evidence library
HarnessPro
改了什么：详见证据卡
HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs
无标签：用模型生成的测试评价候选，不访问标准答案或官方测试；按性能与成本的不同偏好产出 lite、medium、heavy 三档。
Label-free: evaluates candidates with model-generated tests, without accessing standard answers or official tests; produces lite, medium, and heavy tiers for different performance–cost preferences.
HF 推荐： HF recommendations: 2026-10-20 版本更新： Version update: 2026-10-20
其余 4 项待核 4 other checks await verification 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 有代码链接：待核 Code link: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
二、改进器自身进化 2. Evolving the improver (meta / multi-level) 2
原始发表 Published 2026-10-26
二、改进器自身进化 2. Evolving the improver (meta / multi-level)
Skill^∞
改了什么：元技能与任务技能
Skill^{infty}: Recursive Self-Improvement for Agent Skill Optimization
元技能用同一流程优化任务技能并优化自身
A meta-skill uses the same optimization loop to improve task skills and itself.
HF 推荐： HF recommendations: 2026-10-27
其余 4 项待核 4 other checks await verification 有代码链接：待核 Code link: Not verified 跨模型：待核 Cross-model: Not verified 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-09
二、改进器自身进化 2. Evolving the improver (meta / multi-level) 已收录证据库 In the evidence library
Harness²
改了什么：详见证据卡
Harness^2: Recursive Agent Harnessing for an Open World
开放世界、无真值验证器：任务层用「提出—试探—组合」对比步引导编辑，上层把编辑经验沉淀到持续改进 Harness。
Open world without a ground-truth verifier: the task level guides edits through “propose–probe–compose” contrastive steps; the upper level distills editing experience into a continually improving harness.
HF 推荐： HF recommendations: 2026-10-14
其余 4 项待核 4 other checks await verification 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 有代码链接：待核 Code link: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces) 7
原始发表 Published 2026-10-20
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces) 已收录证据库 In the evidence library
Evolving Agent Interfaces (Agent S4)
改了什么：详见证据卡
Evolving Agent Interfaces
元 Agent 阅读合成任务上的无成功标签轨迹，写入并修订接口仓库；种群选择按元 Agent 看不到的验证任务决定保留哪些仓库。
A meta agent reads trajectories without success labels on synthetic tasks, writing and revising interface repositories; population selection retains repositories using validation tasks hidden from the meta agent.
HF 推荐： HF recommendations: 2026-10-21
✓ 留出集 Held-out 来源 Source papers/agent-s4.md ✓ 有代码链接 Code link 来源 Source papers/agent-s4.md 其余 2 项待核 2 other checks await verification 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
原始发表 Published 2026-10-19
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces)
TEAL
改了什么：工具库与用法技能
Unsupervised Tool Evolution for Coding Agents in Data-Lake Analytics
无标签设置下跨问题演化工具，只保留 Worker 复跑时采用且有证据支持的修改
Tools evolve across unlabeled problems; changes are retained only when the Worker adopts them on replay and the answer has supporting trajectory evidence.
HF 推荐： HF recommendations: 2026-10-21
✓ 同预算 Matched budget 来源 Source lifting Sonnet-5 at 15 steps by 9.3 points 其余 3 项待核 3 other checks await verification 有代码链接：待核 Code link: Not verified 跨模型：待核 Cross-model: Not verified 留出集：待核 Held-out: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-19
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces)
GRSE
改了什么：多 Agent 技能库
Which Agent's Skill Should Change? Group-Relative Credit Assignment for Evolving Multi-Agent Teams
不更新权重，按组内相对比较把团队奖励归因到具体技能再修订
Relative within-group comparisons attribute team rewards to individual skills for revision without updating model weights.
HF 推荐： HF recommendations: 2026-10-22
✓ 有代码链接 Code link 来源 Source https://github.com/yananma1029/GRSE-Skills 其余 3 项待核 3 other checks await verification 跨模型：待核 Cross-model: Not verified 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-16
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces)
HexaShell
改了什么：机器人感知工具代码
HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
每次失败后读日志写或修工具，冻结后在新布局上测试
After failures, the agent reads logs and writes or repairs perception tools; the tools are frozen before testing on new layouts.
HF 推荐： HF recommendations: 2026-10-19 版本更新： Version update: 2026-10-20
✓ 有代码链接 Code link 来源 Source https://github.com/HexaFuture/HexaShell ✓ 留出集 Held-out 来源 Source test them on ten new layouts of the same seed 其余 2 项待核 2 other checks await verification 跨模型：待核 Cross-model: Not verified 同预算：待核 Matched budget: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-15
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces)
TokenEconomizer
改了什么：注入的微技能与工具集
TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents
跨任务序列根据经验注入技能、裁剪长期无用工具，作用于 Claude Code/Codex/Copilot CLI
Across tasks, the controller injects experience-driven micro-skills and removes persistently unused tools for Claude Code, Codex, and Copilot CLI.
HF 推荐： HF recommendations: 2026-10-19
✓ 有代码链接 Code link 来源 Source https://github.com/microsoft/TokenEconomizer ✓ 跨模型 Cross-model 来源 Source Across Claude Code, Codex, and GitHub Copilot CLI ✓ 留出集 Held-out 来源 Source generalize to diverse domains 其余 1 项待核 1 other checks await verification 同预算：待核 Matched budget: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-13
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces)
Socratic Skill Learning
改了什么：可迁移的 Skill 文件
Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills
Student 通过实践编写并修订 Skill，在新会话的留出用例上复用
The student writes and revises transferable Skill files through practice, then reuses them on held-out cases in new sessions.
HF 推荐： HF recommendations: 2026-10-22
✓ 有代码链接 Code link 来源 Source https://github.com/NJU-LINK/Socratic-Skill-Learning ✓ 跨模型 Cross-model 来源 Source Across three model pairs on shared families ✓ 留出集 Held-out 来源 Source held-out evaluation in fresh solver sessions 其余 1 项待核 1 other checks await verification 同预算：待核 Matched budget: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
原始发表 Published 2026-10-11
三、组件级进化 3. Component-level evolution (skills, memory, context, interfaces)
ViewCraft
改了什么：证据接口与读取技能
VIEWCRAFT: Evolving What Multimodal Agents See and How They Read It
失败轨迹驱动修订，回放与配对评估决定保留，同 token 预算下仍有优势
Failure traces drive revisions to evidence interfaces and reading skills; replay and paired evaluation select retained changes, with gains under a matched token budget.
HF 推荐： HF recommendations: 2026-10-14
✓ 有代码链接 Code link 来源 Source https://github.com/Little-Fridge/ViewCraft ✓ 跨模型 Cross-model 来源 Source The learned skills also transfer across executors ✓ 同预算 Matched budget 来源 Source The advantage persists at a matched input-token budget. 其余 1 项待核 1 other checks await verification 留出集：待核 Held-out: Not verified 自动判断，未核实；不代表已审阅正文证据。 Automated, unverified; this does not establish full-paper evidence.
四、模型与 Harness 协同进化 4. Model–harness co-evolution 3
原始发表 Published 2026-10-20
四、模型与 Harness 协同进化 4. Model–harness co-evolution 已收录证据库 In the evidence library
HarnessAdapt-RL
改了什么：详见证据卡
Training the Whole Agent, Not Just the Model
RL 期间，一个教师生成任务相关的机制级技能，另一个根据反复出现的交互问题演化工具和 hooks；只保留使通过率或可训练率提高且另一指标不降的修改。
During RL, one teacher generates task-related mechanism-level skills, while another evolves tools and hooks from recurring interaction problems; retains only edits that improve pass rate or trainability without lowering the other metric.
HF 推荐： HF recommendations: 2026-10-21
其余 4 项待核 4 other checks await verification 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 有代码链接：待核 Code link: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
原始发表 Published 2026-10-18
四、模型与 Harness 协同进化 4. Model–harness co-evolution 已收录证据库 In the evidence library
CoEvo-RSI
改了什么：详见证据卡
CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?
元 Agent 借共享演化档案双向继承历史 Harness 与模型检查点，用自身拒绝采样轨迹微调，缓解「脑体错配」。
A meta agent uses a shared evolutionary archive for bidirectional inheritance of historical harnesses and model checkpoints, fine-tuning on its own rejection-sampled trajectories to mitigate “brain–body mismatch.”
HF 推荐： HF recommendations: 2026-10-21
其余 4 项待核 4 other checks await verification 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 有代码链接：待核 Code link: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
原始发表 Published 2026-10-16
四、模型与 Harness 协同进化 4. Model–harness co-evolution 已收录证据库 In the evidence library
Self-HPO
改了什么：详见证据卡
Self-Evolving Coding Agents via Harness-Policy Optimization
H-step 由当前策略分析自身失败并提出 Harness 候选；P-step 在选定 Harness 下做 RL；仍未解决的任务成为下一轮搜索集。
In the H-step, the current policy analyzes its own failures and proposes harness candidates; in the P-step, it performs RL under the selected harness; still-unsolved tasks become the next round’s search set.
HF 推荐： HF recommendations: 2026-10-20
✓ 有代码链接 Code link 来源 Source papers/self-hpo.md 其余 3 项待核 3 other checks await verification 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.
理论与综述 Theory & surveys 1
原始发表 Published 2026-10-13
已收录证据库 In the evidence library
Self-Evolving Agent Harness (theory)
改了什么：详见证据卡
Self-Evolving Agent Harness
把 Harness 自进化拆成「更新对象」与「提案/准入机制」，给出持续 PAC 学习下的泛化界、保留界与基于新鲜验证数据的准入规则。理论界依赖明确假设。
Decomposes harness self-evolution into update targets and proposal/admission mechanisms, providing generalization bounds, retention bounds, and admission rules based on fresh validation data under continual PAC learning. The theoretical bounds depend on explicit assumptions.
HF 推荐： HF recommendations: 2026-10-16
其余 4 项待核 4 other checks await verification 留出集：待核 Held-out: Not verified 同预算：待核 Matched budget: Not verified 跨模型：待核 Cross-model: Not verified 有代码链接：待核 Code link: Not verified 证据来自人工整理的证据卡；逐项结论以原卡为准。 Curated evidence; see the evidence card for individual findings.